1、spark streaming direct方式读取kafka性能要好很多，缺点是它不会去zookeeper更新offset，这将导致
基于zk的KafkaOffsetMonitor监控失效。由于我们流处理的数据量大并且健壮性要求高，我们需要
通过KafkaOffsetMonitor来实时监控数据消费情况，故自己实现offset推送到zk，如下：

2、spark代码：

object KSConvertStreaming{
val savaLocal = "/xxx/parquet/%s/year=%s/month=%s/day=%s"
def jsonConvert(jsonStrs: Iterator[String]): Iterator[(String, ArrayBuffer[String])] = {
StreamingUtils.init
val typeMap = scala.collection.mutable.Map[String, ArrayBuffer[String]]()
jsonStrs.foreach(
x => {
val res = StreamingUtils.mapToStr(x)
if (null != res) {
val msgType = res.get(0)
if (!typeMap.contains(msgType)) {
typeMap += (msgType -> new ArrayBuffer[String]())
}
typeMap(msgType) += res.get(1)
}
}
)
typeMap.iterator
}
def main(args: Array[String]): Unit = {
val Array(maxPartition, maxNumber,windownsS,groupName,maxRatePerPartition) = args
val topicName = "xxx"
val kafkaAddr = "xxx:9092,xxx:9092,xxx:9092"
val kafkaParams = Map[String, String](
"metadata.broker.list" -> kafkaAddr,
"group.id" -> groupName,
"auto.offset.reset" -> "largest"
)
val topics = Set(topicName)
println(s"maxPartition -------- $maxPartition")
println(s"maxNumber -------- $maxNumber")
println(s"windownsS -------- $windownsS")
println(s"groupName -------- $groupName")
val sparkConf = new SparkConf().setAppName("Streaming_Convert")
.set("spark.yarn.executor.memoryOverhead","1024")
.set("spark.streaming.kafka.maxRatePerPartition",maxRatePerPartition) //此处为每秒每个partition的条数
.set("spark.serializer", "org.apache.spark.serializer.KryoSerializer")
.set("spark.reducer.maxSizeInFlight", "1m")
val sc = new SparkContext(sparkConf)
val sqlContext = new org.apache.spark.sql.SQLContext(sc)
val ssc = new StreamingContext(sc, Seconds(windownsS.toInt)) //秒单位
val topicDirs = new ZKGroupTopicDirs(groupName,topicName)
val zkClient = new ZkClient("xxx:2181,xxx:2181,xxx:2181",Integer.MAX_VALUE,100000,ZKStringSerializer)
val children = zkClient.countChildren(topicDirs.consumerOffsetDir)
var kafkaStream: InputDStream[(String, String)] = null
var fromOffsets: Map[TopicAndPartition, Long] = Map()
val messageHandler = (mmd: MessageAndMetadata[String, String]) => (mmd.topic, mmd.message())
if (children > 0) {
for (i <- 0 until children) {
val partitionOffset = zkClient.readData[String](s"${topicDirs.consumerOffsetDir}/$i")
val tp = TopicAndPartition(topicName, i)
fromOffsets += (tp -> partitionOffset.toLong)
}
kafkaStream = KafkaUtils.createDirectStream[String, String, StringDecoder, StringDecoder, (String, String)](ssc, kafkaParams, fromOffsets, messageHandler)
}
else {
kafkaStream = KafkaUtils.createDirectStream[String, String, StringDecoder, StringDecoder](ssc, kafkaParams, topics)
}
var offsetRanges = Array[OffsetRange]()
kafkaStream.transform { rdd =>
offsetRanges = rdd.asInstanceOf[HasOffsetRanges].offsetRanges
rdd
}.map(_._2).foreachRDD {
rdd =>
val xRDD = rdd.flatMap(_.split("\n")).mapPartitions(x => jsonConvert(x))
xRDD.persist(StorageLevel.MEMORY_ONLY)
val typeCountMap = xRDD.map(x => (x._1, x._2.size)).collect()
val typeMap = scala.collection.mutable.Map[String, Long]()
typeCountMap.foreach(x => {
if (!typeMap.contains(x._1)) {
typeMap(x._1) = x._2
} else {
val tmpCount = typeMap(x._1)
typeMap(x._1) = x._2 + tmpCount
}
})
var totalCount: Long = 0
typeMap.foreach(x => {
println(s"${x._1}:${x._2}")
totalCount += x._2
})
println(s"total : $totalCount" )
val sortedMap = collection.mutable.LinkedHashMap(typeMap.toSeq.sortWith(_._2 > _._2): _*)
sortedMap.foreach {
x => {
val pointType = x._1
val count = x._2
println(s"save type( $pointType ) count( $count )")
val jsonRDD = xRDD.filter(x => x._1 == pointType).
flatMap(x => x._2)
var partitionNum = count / maxNumber.toLong
if (partitionNum == 0) {
partitionNum = 1
}
if (partitionNum > maxPartition.toLong) {
partitionNum = maxPartition.toLong
}
println(s"\trepartition ( $partitionNum )")
val arrType = pointType.split('-')
sqlContext.read.json(jsonRDD).
repartition(partitionNum.toInt).
write.mode(SaveMode.Append).
parquet(savaLocal.format(arrType(0),arrType(1),arrType(2),arrType(3)))
}
}
xRDD.unpersist()
for (o <- offsetRanges) {
val zkPath = s"${topicDirs.consumerOffsetDir}/${o.partition}"
ZkUtils.updatePersistentPath(zkClient, zkPath, o.untilOffset.toString)
}
}
ssc.start()
ssc.awaitTermination()
}
}

ITPub博客

spark direct kafka 将offset推到zookeeper

最新文章